Foresight cannot be reconstructed from the outcome
Once an outcome is known, the route leading to it can feel coherent and almost inevitable. Earlier uncertainty, plausible alternatives and genuine surprise become harder to recall. Classic work by Fischhoff and Beyth showed that remembered probabilities could shift towards outcomes participants later believed had occurred. The practical lesson is not that every researcher will respond identically, but that memory alone is a weak record of prior expectation.
A dissertation can externalise that record. A dated proposal, hypothesis sheet, analysis plan and decision diary preserve what was considered possible before the relevant result appeared. The later report can then distinguish what was supported, contradicted or first suggested by the data. The aim is accurate chronology, not a claim that planning eliminates every source of bias.
Notice where a polished narrative can rewrite the process
In the introduction, an unexpected association may be recast as the only logical hypothesis. In the method, discarded alternatives disappear and the chosen analysis looks self-evident. In the discussion, theories that fit receive attention while contrary expectations fade. In qualitative research, themes that dominated the final coding may be described as though their final form had been fully anticipated before open engagement with the material.
A coherent report is desirable. The difficulty is confusing the order used to explain the study with the order in which knowledge was acquired. Readers need to know which proposition guided the design and which emerged later. Post-result reasoning can generate valuable research, but it should be labelled exploratory and should not be presented as independently confirmed by the same observations that inspired it.
Before and after: an unexpected near-null relationship
Before analysis: “Based on the reviewed studies, I expect a moderate positive association between study time and test performance. A weaker association is plausible because self-reported time may be imprecise; a curved pattern is another possibility. The primary analysis is the pre-specified linear model, with alternative forms treated as exploratory.” This records the preferred prediction, alternatives, a measurement concern and the analysis status.
After analysis: The estimated linear association is small and uncertain. It would now be misleading to claim that measurement problems made a null result obvious all along. A transparent report says: “The result did not match the primary expectation. Limited measurement accuracy is one possible explanation generated during interpretation; the present data do not independently test it.” Any subsequently examined curved pattern is labelled exploratory.
The example contains no universal threshold. Suitable models, uncertainty measures and interpretations depend on the field, design and data. Its purpose is chronological: prediction, observation and newly generated explanation remain separate objects.
Keep surprise scientifically useful
An unexpected result is not automatically an error, and exploratory analysis is not inferior research. New questions often arise because observations resist an existing account. The integrity problem begins when the discovery is retold as a prediction. Name the surprise, check data and model assumptions, consider alternative explanations and derive a testable next hypothesis.
Ideally, a newly generated hypothesis is assessed on independent data. Where a dissertation cannot collect another sample, state that limitation and propose a specific future test. Avoid repeatedly dividing the existing data until a persuasive narrative appears. Additional analyses, their timing and their trigger belong in the analysis record.
Write a discussion that preserves uncertainty
First relate the result to the question as it was originally framed. Then compare it with the documented expectation. Separate explanations derived in advance from interpretations prompted by the result. Offer plausible alternatives rather than presenting the realised outcome as inevitable, and explain what evidence would distinguish them.
Words such as “obviously”, “inevitably” and “as expected” deserve a check against the frozen plan. Literature found after analysis may legitimately inform interpretation, but its later role should be clear. It can support a proposed explanation; it cannot demonstrate that the explanation was your prior prediction.
Useful safeguards short of formal preregistration
Preregistration can provide a time-stamped study plan, but it may not be required or suitable for every assignment. A dated proposal, message to a supervisor, versioned hypothesis sheet or frozen analysis note can still create a comparison point. The essential feature is that the record predates knowledge of the relevant outcomes and that later amendments remain visible.
Classify each analysis as confirmatory, exploratory or a sensitivity check. When a plan must change for a sound reason, record the reason and whether the results were already known. Transparency should not prevent sensible adaptation; it should preserve when and why adaptation happened.
A hindsight audit before submission
- Compare hypotheses and methods line by line with the earliest dated plan.
- List each departure with its date, information state and rationale.
- Check whether unexpected findings are recognisable as unexpected in the prose.
- Separate new explanations from tests specified before analysis.
- Replace deterministic hindsight with language that retains alternatives and uncertainty.